Poster #119 - Detecting Developmental Trends and Gender Differences in Emotion Expression in Children’s Poetry Using Sentiment Analysis

• In Event: 3-106 - Poster Session 11
  In Poster Session: PS 11 Section - Methods, History, Theory

Integrative Statement

The ubiquity of online textual data offers potential for researchers to explore central questions about child development. As an exemplar, this study employed a technique for analysing large compendiums of text to explore how emotion expression changes over time. Sentiment analysis is a way of automatically detecting the valence (positive or negative) of text (Bing, 2015). This technique identifies affective properties of text efficiently and with relative accuracy, and thus may be leveraged to investigate developmental changes in affect. Research suggests that positive affect decreases throughout development, especially during early-middle adolescence in concordance with disruptive life changes and shifting self-perceptions (Larson et al., 2002). Research also suggests that these trajectories differ among boys and girls, though findings are mixed as to whether boys or girls experience sharper declines in affect (Chaplin & Aldao, 2013). Testing these assertions from a different perspective, the purpose of this study was to use sentiment analysis to explore developmental changes and gender differences in valence of over 35,000 poems written by children and adolescents.
N=35,546 poems were obtained using web scraping of an online, open-publishing children’s poetry website (permissions obtained from domain host). The average poem length was 18.55 words (SD=15.76). Poems were sorted by school grade (grades 1-10) and included the first name of the author (used to infer gender by matching with a widely used list of common male and female names, Kantrowitz & Ross, 1994). This resulted in 20,595 poems identified as written by girls and 14,951 by boys. Each poem was separated into individual words (i.e., tokenized), and function words (e.g., ‘the’) were omitted. Valence was assigned to each word using the Valence ratings of the Valence, Arousal, Dominance lexicon (Mohammad, 2018), which has ratings of over 20,000 English words identified via crowdsourcing (0=highly negative, 1=highly positive). For example, the word ‘heart’ registers as .82 and the word ‘death’ registers as .03. For each poem we calculated the average valence of the words within that poem.
Differences in the content of poems were tested by comparing commonly used words across age groups (see Figure 1). Among the results, younger children mentioned concrete objects (e.g., ‘cat’, ‘dog’) more frequently, whereas older children mentioned abstract concepts (e.g., ‘love’, ‘feel’) more often. Overall valence decreased significantly from Grade 1 to 10, β=-.002, t(35544)=-5.93, p<.001, while controlling for the number of total words in each poem. There was also a significant gender difference, with boys’ poems (M=.62, SD=.11) having significantly lower valence score compared to girls’ poems (M=.64, SD=.11; t(1,35544)=18.62, p<.001). Finally, there was a significant Grade x Gender interaction, β=.002, t(35541)=3.01, p=.003, with the decrease in positive valence being significantly negative for girls only (see Figure 2). Overall, results are consistent with the notion that affect decreases with age during childhood and adolescence. Results further suggest that this decrease is particularly strong among girls. Sentiment analysis represents a novel way of exploring developmental trends in emotion expression. Used in conjunction with traditional methods, sentiment analysis offers additional evidence for diverse psychological research questions.

Authors

• Will E Hipson, Carleton University
  Presenting Author

• Saif M Mohammad, National Research Council Canada
  Non-Presenting Author